Enabling Active Storage on Parallel I/O Software Stacks. Seung Woo Son Mathematics and Computer Science Division

Transcription

1 Enabling Active Storage on Parallel I/O Software Stacks Seung Woo Son Mathematics and Computer Science Division MSST 2010, Incline Village, NV May 7, 2010

2 Performing analysis on large data sets is often frustrating Original Data Acquire data or simulate Target Data Data integration & selection Preprocessed Data S3D simulations for combustion research are producing TB of data per simulation Preprocessing (pattern recognition & feature extraction) Patterns Model construction Knowledge Interpretation (interactive analysis & visualization) Scientists and engineers spent too much time on data manipulation, especially moving and reorganizing data 2

3 Talk outline Motivation Active storage in parallel file systems Our prototype Enhanced runtime interface that uses embedded analysis kernels Runtime stripe alignment Server-to-server communication for reduction and aggregation Experimental evaluation Conclusion 3

4 Active storage in parallel file systems client nodes C 1 C 2 C m filter Interconnect network filter File read filter Reduced data sent to client Library or user space implementation; not well integrated into I/O software stacks Targeting applications that manipulates fundamentallyindependent data sets server nodes S 1 S 2 Active storage is a technique for performing data transformations in the storage system S n Lack of reduction and aggregation on the storage nodes E. Riedel et al., Active disks for large-scale data processing, IEEE Computer, J. Piernas et al., Evaluation of active storage strategies for the Lustre parallel file systems, in SC,

5 We enable active storage on parallel I/O software stack 1. Enhanced runtime interface (API) to enable active storage operations 2. Runtime data stripe alignment 3. Server-to-server communication primitives for complex analysis 5

6 Enhanced runtime I/O interface to trigger embedded analysis kernels sum = 0.0; MPI_File_open(&fh); double *tmp = (double*)malloc(nitem*sizeof(doble)); offset = rank * nitem * type_size; MPI_File_read_at(fh, offset, tmp, nitem, MPI_DOUBLE, &status) for(i=0; i<nitem; i++) sum += tmp[i] Conventional MPI-based... MPI_File_open(&fh);... MPI_File_read_ex(, SUM, )... <client> Active storage based for(i=0; i<nitem; i++) sum += tmp[i]; <server> 6

7 Why MPI? MPI is a widely used interface There are a large number of applications Therefore, it might be relatively easy to migrate MPI specification provides interfaces where user functions can be embedded into it Enabling the incorporation of data mining and statistical functions easily Hint mechanism Passing kernel specific argument to the server, e.g., data types 7

8 Mapping embedded analysis kernels into I/O pipeline pvfs state machine machine pvfs_pipeline_sm { state fetch static int fetch_data { disk I/O; { run fetch_data; normal_op => dispatch; active_op => do_comp; } static int dispatch_data { send the data; } } } state do_comp { run do_comp_op; success => dispatch; } state dispatch { run dispatch_data; success => check_done; } state check_done { run check_done_action; not_done => fetch; default => terminate; } static int do_comp_op { for(i=0;i<nitem;i++) } sum += tmp[i] ; 8

9 Computational unit is often not perfectly aligned to file stripe unit n-dimensional data set 80 bytes day1 day2 day bytes bytes bytes 9

10 I/O pipeline with data alignment 10

11 Server-to-server communication for reduction and aggregation 1. Randomly choose initial centers 2. Assign each point to the nearest center Reduction and aggregation can be done on client side (e.g., simple statistical operations) 3. Update centers (mean of members) 4. Repeat until convergence Complex analysis kernels (e.g., k-means clustering) requires broadcast and reduction during iterative execution K-means cluster algorithm 11

12 K-means clustering is performed purely on the server side! 12

13 Benchmarks and evaluation platform Name description Base (sec) Input data % of filtering sum Global reduction MB ~100% grep kmeans vren String pattern matching 1.49 K-means clustering algorithm 0.44 Parallel volume rendering Test cluster MB (4M of 128 string) 40 MB (1M*10 dim of double) 103MB (300*300*300 of float) 32 nodes Dual Intel Xeon Quad Core 2.66 MHz Main memory Storage capacity Interconnection network GPU accelerator 16GB ~200GB per node 1 Gb Ethernet 2 NVIDIA C1060 GPU card ~100% 90% 97% 13

14 All benchmarks are I/O dominant 64.4% time is spent on I/O Benchmarks are executed using 4 nodes 14

15 Moving computation to storage server (AS) improves performance significantly TS: Traditional Storage, 4 client nodes and 4 server nodes AS: Active Storage, 4 server nodes 15

16 Our approach is scalable w.r.t the different number of nodes SUM benchmark Fixed data size: 512MB 16

17 Putting client and server together SUM benchmark using 1 node No Inter-node communication, but Inter-process communication still exists To achieve this in reality, client should be aware of storage layout! 17

18 Conclusion Target Data Preprocessed Data Model construction Patterns Knowledge Interpretation Original Data Acquire data or simulate Data integration & selection Preprocessing Enabling active storage through: 1. Enhanced runtime interfaces (APIs) 2. Runtime stripe alignment 3. Server-to-server aggregation Enabling Active storage within parallel I/O software stack removes not only internode data transfer, but also inter-process data communication, resulting in a huge performance improvement for data-intensive analysis applications 18

19 Acknowledgments Department of Energy for funding this work Phil Carns, Sam Lang, Rob Ross, Rajeev Thakur (ANL) Alok Choudhary, Prabhat Kumar, Wei-Keng Liao, Berkin Ozisikyilmaz (NWU) 19

20 Thanks! 20

21 Future work Function shipping More flexible hint mechanism Hadoop style execution Write output result to the local storage Scalability analysis NCSA Lincoln cluster: 192 compute nodes and 96 NVIDIA Tesla S1070 accelerator units. More benchmarks/applications Visualization and Bioinformatics 21

22 Give hints to file servers for more information MPI_Info info; MPI_Init(); MPI_Comm_rank(); MPI_Info_create (&info); MPI_Info_set (info, key, val ); MPI_File_open (, info, ); MPI_Info_free (&info); MPI_Finalize(); <general MPI hint mechanism> Data type and operators are sufficient for simple operations, e.g., sum Some kernels might need more information to perform correct computation Grep: string length per line (128), search pattern ( aaaaa ) K-means: number of dimension (10), number of clusters (20), threshold value (0.001), etc. 22

23 Our approach is scalable w.r.t the different data set sizes SUM benchmark Fixed number of nodes: 4 23

24 Data mining kernels can be compute intensive 1. Randomly choose initial centers 2. Assign each point to the nearest center 3. Update centers (mean of members) 4. Repeat until convergence K-means clustering algorithm 24

25 Our approach is scalable w.r.t number of nodes to execute and data set size Fixed data set size = 1M data points Delta = AS+GPU: active storage with GPU Fixed # of nodes = 4 Delta =